Low-Power Semantic Segmentation on Embedded Systems for Traffic in Asian Countries

Abstract

In the real-world road scene, it needs a quick and powerful system to detect any traffic situation and compute prediction results in a limited hardware environment. In this paper, we present a semantic segmentation model that can perform well in the complex and dynamic Asian road scenes and evaluate the accuracy, power, and speed of MediaTek Dimensity 9000 platform. We use a model called Deep Dual-resolution Networks (DDRNets) in PyTorch, and deploy TensorFlow Lite format in MediaTek chip to assess our model. We choose the two-stage training strategy and utilize the decreasing training resolution technique to further improve results. Our team is the first-place winner in Low-power Deep Learning Semantic Segmentation Model Compression Competition for Traffic Scene in Asian Countries at IEEE International Conference on Multimedia & Expo (ICME) 2022.